Detecting Chickens From Social Media Data

Brad Cooley, Arpan Ojha School of Informatics, Computing, and Engineering at Indiana University

Myrna Cadena, School of Veterinary Medicine, University of California, Davis

Method

Top 1

Accuracy

Train

Accuracy

Time to

train/detect(s)

Neural Networks

CNN(Baseline)

75 95.8 240/60

VGG

-19 87 89 28800/720

ResNET

-50 98 99 3600/300

ConvNext

94 99 60/30

Transformers

ViT

35 45 600/5

ConvNeXt SOTA performance on farm vs game fowl. Green border indicates

correct classification whereas red border indicates incorrect classification.

Train/Test images obtained from Kaggle, Reddit, and Google.

1. Motivation

2. Challenges

Using ConvNeXt and transfer learning, our implementation achieves

breed-specific labeling and, when not available, differentiation between

game fowl and domestic fowl.

3. Initial Modeling of Chicken Body Shape

4. Experiments

▪ We experimented with four different models:

▪ Baseline CNN, horrible classification when

other animals were introduced (10-15%).

▪ ResNet50, decent classification, but not good

enough to be breed specific

▪ ConvNeXt, good breed specific classification

and excellent species classification(90% top5)

▪ ViT, mediocre species classification(~35%)

5. Results

Above: Model complexity. All

images use adam and cross

entropy loss

Right: Comparison to various

models in terms of top 1 test ,

training accuracy and time to

detect .

There are times that our model didn't classify images even remotely correct.

For example, the middle image was classified as a lapdog.

6. Future Work

▪ Refinement of ConvNeXt labeling techniques

▪ Applying ConvNeXt to video as compared to images

▪ More broad data collection strategies and sources

▪ Try DeiT and other modern transformers.

Visual depiction of game

versus farm fowl

▪ Given two broad

categories of chicken

game and farm

▪ Estimate social media

websites for chicken data

▪ Also, jointly estimate

type of chicken, game or

farm based for diseased

protection and isolation

▪ Need to identify contours and feather data properly

▪ Deeper layers perform better with more expressive

details documented

▪ Model needs to be as discriminative as possible, so

we train against similar looking animals (e.g.

squirrels)

▪ Handle different species, color, and orientation of

chickens including texture of feathers.

▪ Solve limited data issues with stitching,

translation, and warping

▪ Model needs to be of the "chickens versus the

world" type, where chickens are classified against

other animals / observable items in the world.

▪ Out-of-the-box computer vision models (ResNet10,

ResNet50, etc.) tend to have a hard time classifying

different breeds of the same species

▪ This problem becomes increasingly more

pronounced with data scraped from various social

media sources (Twitter, Craigslist, etc.)

▪ We propose a way to use state-of-the-art (SOTA)

computer vision models to accurately label different

breeds of the same species (in our case, Chickens)

▪ We utilize ConvNeXt and transfer learning to achieve

breed specific classification where possible

Acknowledgements: This project is in collaboration with the CE Poultry Lab at the University of California – Davis.

Method

Layers/epochs Parameters (Million) Optimizer / Loss Function

CNN (baseline)

3/25 0.001 adam/Binary Cross entropy

VGG

-19 19/20 138 adam/Cross entropy

ResNET

-50 50/9 23 adam/Cross entropy

ConvNext

-/2(90 actual) 350 adamw/Cross entropy

ViT

12/20 87 adam/Cross entropy

Based on results, ConvNeXt is best suited for our purpose. It has the lowest

false positives and negatives. Test detection time needs to reduced.